Reliability of transfer learning across tasks in molecular generative models

Determine whether molecular generative models trained jointly on multiple tasks relevant to structure-based drug design (such as pocket-conditioned de novo ligand design and molecular docking) can reliably leverage transfer learning to improve performance across tasks, and identify conditions under which such cross-task transfer is consistently beneficial.

Background

OMTRA is introduced as a unified multi-task generative model for structure-based drug design, capable of handling ligands, protein pockets, and pharmacophores under a multi-modal flow matching framework. The work evaluates the effects of protein-free pretraining and multi-task training and finds the benefits to be modest and inconsistent.

These observations motivate a broader uncertainty about whether molecular generative models can consistently extract and apply transferable knowledge across tasks. Resolving this would clarify the practical value of multi-task training and guide future architectural and training decisions for models like OMTRA.